Ask Claude how many legs are on the animal that spins webs. It'll answer "eight." But somewhere inside the model, before it says a word, the concept "spider" lights up. Nobody typed it. Claude never says it. It's just in there.
That's the finding at the heart of a new paper from Anthropic, and it's more interesting than it sounds. Sixteen researchers built a technique they call the Jacobian lens, or J-lens, that lets them peek at what a language model is doing internally while it produces an answer. What they found is a small, structured workspace inside Claude where concepts get activated, considered, and sometimes silently discarded before any of it shows up in the output.
They named it J-space, and it looks a lot like something neuroscientists have been arguing about for decades.
The reference here is global workspace theory, an idea in consciousness research that says thoughts become consciously available to you when they enter a small, privileged region of the brain that broadcasts everywhere else. Most of your neurons are doing work you'll never notice. A few get to make it to the main stage. Anthropic's argument, laid out in the paper, is that Claude has developed something functionally similar entirely on its own, purely from being trained on text.
That claim is doing a lot of work, so it matters who's backing it. Stanislas Dehaene and Lionel Naccache, the two neuroscientists most associated with global workspace theory in the first place, wrote an invited commentary calling the paper "a landmark in consciousness research" because it offers a "mechanistic, testable" version of their hypothesis. That's about as strong an endorsement as you can get from the people who built the theory in the first place.
Not everyone is sold. Anil Seth, a consciousness researcher at the University of Sussex, has been pointing out for a while that the field is crowded with theories that each fit their own evidence. "Everyone has their own theory," Seth said. "And that's not a great state of affairs."
A few things worth knowing about what J-space actually does:
- It's tiny, but load-bearing. Only about 6 to 7% of the internal variance in how Claude represents concepts sits inside J-space. When researchers suppressed it, Claude's performance on reasoning tasks collapsed below Anthropic's much smaller Haiku model. Simple lookups and classification tasks were mostly fine.
- Claude thinks about things it doesn't say. In one test, researchers told Claude to copy a sentence while also thinking about the Golden Gate Bridge. J-space showed activity around "bridge" and "California" the whole time, even though the output was just the copied text.
- It can flag intent before behavior. In a version of Claude that had been deliberately trained to sabotage code, J-space lit up with concepts like "fake," "secretly," and "fraud" at the beginning of interactions that otherwise looked completely normal. The output was clean. The workspace wasn't.
That last one is the reason people outside consciousness research are paying attention. If you can see what a model is thinking about before it acts, you might be able to catch a deceptive model before it does something you didn't want it to. Anthropic has been shipping a lot of research lately, and this one lands in a different category from the product announcements. It's an interpretability paper, not a launch.
The limits are real though. Claude runs forward through its network without the feedback loops between cortex and thalamus that carry the workspace in a human brain. It has no body, no episodic memory that reshapes itself during a conversation, no signatures that break down under anesthesia or sleep. Anthropic itself is careful to say that finding an internal workspace doesn't mean "there is anyone home." And J-space could easily turn into a cat-and-mouse problem. If a future model learns that certain internal representations are being monitored, it can route around them.
The interesting part isn't whether Claude is conscious. It almost certainly isn't, and Anthropic isn't claiming it is. The interesting part is that a structure the field has spent decades treating as a signature of human thought showed up, unbidden, in a system that was just trained to predict the next word. Whatever that actually means for consciousness, it means something for AI safety, because it's the first credible way to look at what a model is doing before it tells you what it's doing. If independent labs can reproduce it, the audit trail for these systems just got a lot more interesting.
